This study investigates the use of machine learning–based segmentation for estimatingCredit Conversion Factors (CCF) within credit risk modeling, with a comparative analysis of household and corporate loan portfolios. CCF is a key parameter in the estimation of Exposure at Default (EAD), reflecting the proportion of undrawn credit expected tobe utilized prior to default. Given the complexity and non-linear nature of borrower behavior, this thesis explores whether data-driven methods can improve the understanding and prediction of CCF.
The methodology combines supervised and unsupervised learning techniques. First, anXGBoost model is employed to identify the most important features driving CCF vari-ation, using SHAP values for interpretability. These features are then used as input forclustering algorithms, primarily K-means, to segment customers into groups with similarrisk profiles. Separate predictive models are subsequently trained for each cluster andevaluated against a global benchmark model using RMSE and MAE.
The results show that a limited number of features explain most of the variation in CCF. For household customers, key drivers include income, on-balance exposure, overdraft be-havior and customer type, while for corporate customers, income, exposure, off-balance commitments, relationship length and interest coverage are most influential. The clustering approach successfully identifies distinct segments with different risk characteristics and underlying drivers.
In terms of predictive performance, clustering provides a marginal improvement over a global model for household data, while results for corporate data remain comparable or slightly weaker. However, clustering significantly enhances interpretability by revealing segment-specific relationships and heterogeneity in risk drivers. Overall, the study demonstrates that machine learning–based segmentation is a valuable tool for understanding credit risk behavior and improving transparency in CCF modeling, even if gains in predictive accuracy are limited.
2026. , s. 57